EPIC: populace-am — Armenia: closing the EG DNA macro–micro gap on public data

[MaxGhenis]

Charter

Build populace-am: Armenia as the third greenfield country on the declarative country-spec lane, following Belgium (#259) and New Zealand (#343). Origin: the OECD's distributional-national-accounts compiler team flagged Armenia as a country that participates in the EG DNA/EG DHW expert groups but cannot submit results — the survey base won't reconcile to national accounts (survey consumption covers 26.9% of macro household consumption per the Central Bank of Armenia's UNECE GENA 2024 presentation), and no wealth survey or tax tabulations exist. That macro–micro reconciliation gap is this stack's core competency.

Armenia is unusually well suited to the lane:

• Public base microdata. ArmStat publishes anonymized ILCS household + member microdata for every year 2004–2024, free direct download (armstat.am/en/?nid=205), ~5,184 households/yr, income + consumption. LFS member microdata 2014–2024 alongside. Unlike BE-SILC, no licence restriction — releases can be fully open.
• Rich admin targets. SNA-2008 institutional sector accounts with a household-sector chapter; GDP by income generation; SRC-register wage/payroll series by industry, sex, marz, and firm size; OECD Revenue Statistics totals (PIT AMD 554.6bn = 5.87% of GDP in 2023); pension and family-benefit caseloads from the Unified Social Service.
• Smallest rules surface of any country yet. Flat 20% PIT on labor income (since 2023), schedular 5/10% rates by income type, exempt benefits, capped funded-pension contribution, health-insurance contribution + stamp duty (Dec 2025), family benefit, mortgage-interest and social-expense refunds.
• External oracles exist. CEQ WP 43 (Younger & Khachatryan 2017, ILCS 2011), World Bank PIT-reform microsimulation (2019), World Bank fiscal-incidence report (2025) — validation bands, BE-style (no incumbent, no parity gates).
• Timing. Universal income declaration phases in fully by 2026 (everyone files, SRC pre-filled). Person-level tax microdata matures 2026–28; a working frame positions for a survey+tax top-corrected series later.

Lanes

1. v1 demonstrator (NZ recipe, this epic's first milestone): reweight the existing US support pool against ~10–16 public Armenian margins (population age×sex×marz from the 2022 census, revenue totals, wage distribution, benefit caseloads). Engine-free at first — direct-column targets only; support records are never presented as Armenian microdata.
2. v2 native base (BE recipe): ILCS 2024 household spine as the silc_load analogue (income-reference offset, engine column mapping), QRF income repair, clone-and-assign geography over communities constrained to observed marzes, calibration to household-sector accounts + revenue totals.
3. Rules leg: rulespec-am through the existing Axiom adapter (rulespec-nz/be precedents; no policyengine-am). Repo creation is a maintainer decision, tracked separately. The release contract (engine-native H5, formula_owned_export, reform_validation) waits on it; the calibrated frame does not.

Worklist

• am/ country package: country_package.json, source_stages.json, geography_spine.json, target_references.json, gates.json, release_contract.json — BE greenfield gate posture (aggregate_admin + per_family_fit + target_profile_coverage + macro_realism band; no parity gates)
• ledger-am fact packages in Chronicle: value harvest from ArmStat statbank (PxWeb), OECD Revenue Statistics, MoF budget execution — every fact with source URL; census-2022 vintage for the demographic spine
• Fact-harvest worklist doc: exact tables/queries per target reference
• v1 calibration run + diagnostics (solver lessons from BE apply: destination-scale weight init; totals not component+total double counts)
• External-oracle validation note vs CEQ/WB numbers
• rulespec-am decision + PIT/benefit slice (separate issue once decided)

Non-goals / honesty constraints

• No wealth results — Armenia has no wealth survey; a donor imputation would carry the same borrowed-distribution critique as WID's current series. Income/consumption/saving only.
• Calibration hits sum targets; mapping to full SNA/DNA concepts (institutional households, imputed rent, FISIM) is documented analyst work, not claimed machinery.
• Nothing announced or presented externally without the certified-claims gate; complete ≠ certified.

[MaxGhenis]

Rules leg now exists: TheAxiomFoundation/rulespec-am (bootstrap scaffold, content-free until the corpus-first chain lands — sequence in TheAxiomFoundation/rulespec-am#1).

[MaxGhenis]

Diagnostic calibration result (v1 pool reweight — DIAGNOSTIC, not a release; donor records are US support records, never Armenian microdata):

• Donor: the registry-pinned populace_us_2024.h5 (HF snapshot f09f2f3b, sha256 48b9d479…) — its support pool is 57,240 households, matching the BE-pilot vintage, which resolves the package's donor-provenance open question for the diagnostic. Cloned ×11 marzes → 629,640 records.
• Margins: 616 cells — population age × sex × marz (2024-01-01 census-recalibrated cube, 418), ILCS 2024 households by size × marz (66, kept at its own 887,212 total, unrescaled), LFS 2024 employed by age × sex × marz (132). Industry excluded (no deterministic donor→NACE bridge); SRC register, caseloads, consumption excluded per the harvest contract.
• Fit: median |rel. error| 0.079%; 593/616 within 1%, 606/616 within 10%; family totals population −0.0017%, households −1.12%, employment −0.296%. Worst cells are Vayots Dzor small cells (+64% size-2 households, +60% female 15–24 employed) — the smallest marz stressing a 57k pool, as expected.
• Weight health: Kish ESS 28,178 (4.5% of clones), min/max ratio 20,573, zero negative/zero/nonfinite weights. Solved with destination-scale initialization and totals-not-components (both BE solver lessons).
• Independently reconstructed: all 616 achieved values reproduce within 4.9e-8; verification manifest PASS.

Artifacts: ~/PolicyEngine/_local_runs/armenia-814-diagnostic/ (DIAGNOSTIC.md, per-cell fit CSV, weights, manifests) — local only, nothing published.

Maintainer decisions queued for the ledger-am packaging phase: population vintage (2024-01-01 estimate used here vs census-moment), household total authority (ILCS 887,212 vs census 857,545, ~3% tension reported), SRC payroll representation (the contracted 3-way is not published — finest available decompositions harvested), and the ArmStat licence capture (no licence text exists on the site beyond the attribution footer).

🤖 Generated with Claude Code